123 lines
5.5 KiB
Markdown
123 lines
5.5 KiB
Markdown
# Reported deep-view artifact investigation and fix
|
||
|
||
Date: 2026-09-29
|
||
|
||
## Purpose
|
||
|
||
Investigate the reported URL state that displayed large yellow/black geometric regions and sometimes an almost entirely black frame. The production source before this pass is preserved as `index.before.html`. The fix is intentionally narrow: do not restore previously removed expensive work unless evidence points to it.
|
||
|
||
Reported state:
|
||
|
||
- bits: 273
|
||
- re: `-15958919693715511752069626580195447581643016686947107473855747766272418724092005269`
|
||
- im: `3884297345146583343999140717464554120497735725078805729405733239132420430333375719`
|
||
- span: `35679807704506851074208413824041445358585910697267232113437882789120793`
|
||
- palette: 0
|
||
- cycle: 0.008
|
||
- shift: 0.18
|
||
- base iteration: 350, adaptive enabled, quality high
|
||
|
||
At 1108×1582, the adaptive initial target for this region is approximately 1536 iterations and the renderer takes the `fast-extended` path.
|
||
|
||
## Root cause 1: unsafe series jump
|
||
|
||
The integrated fourth-order series approximation was enabled for the reported view. The fast-reference worker selected:
|
||
|
||
- reference pixel ≈ `(69.5, 922.5)`
|
||
- reference length: 1285
|
||
- series jump: 773
|
||
- fourth-order coefficient components up to approximately `2.35e38`, close to the maximum finite f32 magnitude
|
||
|
||
The series validity test checked truncation terms in unscaled orbit space, but did not bound the *intermediate scaled f32 polynomial products* used by WGSL. At screen-domain deltas away from the selected reference, `a4 * d^4` can overflow before the later `ldexp` rescales the term.
|
||
|
||
The reproducible probe in `experiments/reported-view/analyze.mjs` evaluates the exact worker output from the preserved pre-fix source. On a 32×32 viewport sample grid:
|
||
|
||
- pre-fix: 290 / 1024 series starts produce a non-finite fourth-order intermediate
|
||
- fixed: 0 / 1024, because series acceleration is disabled
|
||
|
||
This is sufficient to generate `RANGE` failures and black unknown pixels. Separate CPU surrogate comparison also showed large escape-iteration deviations in this state, so merely adding an overflow clamp would not make the current jump trustworthy enough for publication.
|
||
|
||
### Fix
|
||
|
||
`SERIES_APPROXIMATION_ENABLED=false` now disables production series construction. The code remains present for future work, but the fast-reference worker emits `jump=0` and zero coefficients.
|
||
|
||
Re-enabling series should require at minimum:
|
||
|
||
1. viewport/reference-relative bounds on all scaled polynomial intermediates,
|
||
2. a propagated truncation/rounding error bound carried into the perturbation path,
|
||
3. a corpus test that checks false escapes, not only absolute approximation error,
|
||
4. strict handling when the selected reference is off-center.
|
||
|
||
## Root cause 2: publishing numerical failures before repair
|
||
|
||
The FAST primary frame was colored and presented immediately after the primary perturbation pass, before `recoverNumericalCostOrdered()` repaired `REFERENCE_END`, `RANGE`, and other numerical failures.
|
||
|
||
`COLOR_WGSL` writes alpha zero for `FIELD_UNKNOWN`. Therefore an early provisional frame can appear:
|
||
|
||
- black when there is no usable history,
|
||
- mixed with reprojected old colors when stable history exists,
|
||
- as large geometric regions reflecting reference validity / numerical-failure domains rather than Mandelbrot geometry.
|
||
|
||
### Fix
|
||
|
||
For `fast-extended`:
|
||
|
||
- primary colorization is deferred (`deferPrimaryColor = deferColor || fastExtended`),
|
||
- unresolved statistics are read,
|
||
- numerical failures are repaired,
|
||
- only then is the first provisional recolor/presentation allowed.
|
||
|
||
Direct rendering keeps its existing provisional behavior.
|
||
|
||
## What was *not* reverted
|
||
|
||
The previous FAST coverage-repair dispatch was not restored. The existing full-frame unknown reduction and final invariant still detect untouched reason-0 pixels, and prior A/B coverage tests found no holes. The reported state produced a concrete series failure mechanism, so restoring an O(pixels) repair pass would treat the symptom without addressing the cause.
|
||
|
||
Other removal-pass optimizations are unchanged.
|
||
|
||
## Verification
|
||
|
||
### Regression
|
||
|
||
`node tests/regression.mjs`
|
||
|
||
Result: PASS, 19 WGSL kernels.
|
||
|
||
Additional regression pins were added for:
|
||
|
||
- `SERIES_APPROXIMATION_ENABLED=false`
|
||
- FAST primary-color deferral via `deferPrimaryColor=deferColor||fastExtended`
|
||
|
||
### Reported-view worker probe
|
||
|
||
`node experiments/reported-view/analyze.mjs`
|
||
|
||
See `results.json`.
|
||
|
||
Key result:
|
||
|
||
| Metric | Before | After |
|
||
|---|---:|---:|
|
||
| reference length | 1285 | 1285 |
|
||
| series jump | 773 | 0 |
|
||
| non-finite series-start samples | 290 / 1024 | 0 / 1024 |
|
||
|
||
### Existing integrated CPU benchmark
|
||
|
||
`node experiments/integrated_bench.mjs`
|
||
|
||
Completed successfully. The series section now reports jump 0 at all tested spans, while periodicity, sparse active state, multi-reference, and CPU fallback tests continue to run.
|
||
|
||
### Browser/WebGPU smoke
|
||
|
||
`tests/browser_smoke.mjs` could not run because its expected external Chromium remote-debug endpoint at `127.0.0.1:9333` was not available in this environment. This is an environment limitation, not a test assertion failure. Actual WebGPU visual confirmation of the supplied URL remains the most important local-device check.
|
||
|
||
## Build summary
|
||
|
||
See `build-summary.json`.
|
||
|
||
- before SHA-256: `e69b91ead027927d977a065d27d9e27629110122df20aa470eab877e23f77ac7`
|
||
- fixed SHA-256: `c53c0dc4cb9fa2d2046aaa12ef03cc79c61a710b0c8db36dea1a87be2d9aedd8`
|
||
- HTML size delta: +396 bytes (comments/guard and publication-order logic)
|
||
|
||
The WGSL kernels themselves were not changed in this pass, so kernel hashes remain pinned to the prior 19-kernel set.
|